perf(skills-page): lazy-fetch the catalog instead of bundling 34MB into JS (#33809)

PR #33748 grew the live skills index from ~2k skills to ~69k, which made
the previous build-time bundling strategy untenable: the skills page's
JS chunk was about to balloon from ~1MB to ~35MB.  Initial page load
on mobile became unusable, search lagged on every keystroke against the
68k-item array, and JSON.parse blocked the main thread at startup.

Three changes:

1. extract-skills.py writes skills.json + skills-meta.json into
   website/static/api/ instead of website/src/data/.  Static-served by
   Vercel as /docs/api/skills.json (gzipped on the wire), same CDN that
   already serves skills-index.json.

2. skills/index.tsx drops the static import and fetches both files in
   parallel on mount.  Loading state shows '…' for the count; failures
   surface a small error pill instead of blanking the page.

3. Search is debounced 150ms and runs against a precomputed lowercase
   haystack stamped onto each row at load time.  Before: array-join +
   toLowerCase per row per keystroke on a 68k array.  After: single
   .includes() per row, deferred until typing settles.

Validation:

| | before | after |
|---|---|---|
| skills.json location | src/data/ (bundled) | static/api/ (CDN) |
| Largest JS chunk | would be ~35MB at 68k skills | 659 KB |
| Initial page render | wait for full parse | immediate, fetch async |
| Per-keystroke filter | join+lowercase x 68k rows | single includes x 68k rows |
| Debounce | none | 150ms |

Built locally for both en and zh-Hans locales; the 34MB skills.json now
lives in build/api/ and is served separately rather than inlined into
the page's bundle.

skills.json and skills-meta.json added to .gitignore — they were already
build artifacts, but the gitignore only listed skills-index.json before.
This commit is contained in:
Teknium 2026-05-28 03:41:43 -07:00 committed by GitHub
parent 6f9182cb34
commit a1eaad2fc0
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 125 additions and 33 deletions

6
.gitignore vendored
View file

@ -78,6 +78,12 @@ mini-swe-agent/
.nix-stamps/
result
website/static/api/skills-index.json
# skills.json + skills-meta.json are build artifacts emitted by
# website/scripts/extract-skills.py during prebuild — keep them out of
# git for the same reason as skills-index.json (large, generated, change
# every build).
website/static/api/skills.json
website/static/api/skills-meta.json
models-dev-upstream/
hermes_cli/tui_dist/*
hermes_cli/scripts/

View file

@ -1,5 +1,5 @@
#!/usr/bin/env python3
"""Extract skill metadata into website/src/data/skills.json for the Skills Hub page.
"""Extract skill metadata into website/static/api/skills.json for the Skills Hub page.
Two data sources:
@ -32,8 +32,12 @@ LOCAL_SKILL_DIRS = [
]
UNIFIED_INDEX_PATH = os.path.join(REPO_ROOT, "website", "static", "api", "skills-index.json")
LEGACY_INDEX_CACHE_DIR = os.path.join(REPO_ROOT, "skills", "index-cache")
OUTPUT = os.path.join(REPO_ROOT, "website", "src", "data", "skills.json")
META_OUTPUT = os.path.join(REPO_ROOT, "website", "src", "data", "skills-meta.json")
# Output to static/api/ so the file is CDN-served at /api/skills.json
# rather than bundled into the page's JS chunk. At 50k+ skills the
# bundled payload was ~26 MB; lazy-fetch keeps the initial page load
# fast and shrinks the JS chunk back to a few hundred KB.
OUTPUT = os.path.join(REPO_ROOT, "website", "static", "api", "skills.json")
META_OUTPUT = os.path.join(REPO_ROOT, "website", "static", "api", "skills-meta.json")
CATEGORY_LABELS = {
"apple": "Apple",
@ -531,7 +535,9 @@ def main():
os.makedirs(os.path.dirname(OUTPUT), exist_ok=True)
with open(OUTPUT, "w", encoding="utf-8") as f:
json.dump(all_skills, f, indent=2)
# Minified — file is served over the wire, not read by humans.
# At 50k+ skills the indented version was ~30% larger.
json.dump(all_skills, f, separators=(",", ":"), ensure_ascii=False)
# Sidecar meta file so the page can render a "Last refreshed" badge
# without changing the shape of skills.json.
@ -547,7 +553,7 @@ def main():
if index_meta:
meta.update(index_meta)
with open(META_OUTPUT, "w", encoding="utf-8") as f:
json.dump(meta, f, indent=2)
json.dump(meta, f, separators=(",", ":"), ensure_ascii=False)
print(f"Extracted {len(all_skills)} skills to {OUTPUT}")
print(f" {len(local)} local ({sum(1 for s in local if s['source'] == 'built-in')} built-in, "

View file

@ -1,7 +1,8 @@
#!/usr/bin/env node
// Runs website/scripts/extract-skills.py and generate-llms-txt.py before
// docusaurus build/start so that:
// - website/src/data/skills.json (imported by src/pages/skills/index.tsx)
// - website/static/api/skills.json (lazy-fetched by src/pages/skills/index.tsx)
// - website/static/api/skills-meta.json (sidecar metadata for the Skills Hub)
// - website/static/llms.txt (agent-friendly short docs index)
// - website/static/llms-full.txt (full docs concat for LLM context)
// all exist without contributors remembering to run Python scripts manually.
@ -30,7 +31,7 @@ const scriptDir = dirname(fileURLToPath(import.meta.url));
const websiteDir = resolve(scriptDir, "..");
const extractScript = join(scriptDir, "extract-skills.py");
const llmsScript = join(scriptDir, "generate-llms-txt.py");
const outputFile = join(websiteDir, "src", "data", "skills.json");
const outputFile = join(websiteDir, "static", "api", "skills.json");
const unifiedIndexFile = join(websiteDir, "static", "api", "skills-index.json");
const UNIFIED_INDEX_URL =
"https://hermes-agent.nousresearch.com/docs/api/skills-index.json";

View file

@ -1,7 +1,5 @@
import React, { useState, useMemo, useCallback, useRef, useEffect } from "react";
import Layout from "@theme/Layout";
import skills from "../../data/skills.json";
import meta from "../../data/skills-meta.json";
import styles from "./styles.module.css";
interface Skill {
@ -21,9 +19,14 @@ interface Skill {
docsPath?: string;
identifier?: string;
installCmd?: string;
/** Lowercase pre-joined haystack used by the search filter.
* Built once at load time so per-keystroke filtering is a single
* `.includes()` per skill instead of array-join + toLowerCase on
* every render. Skipped on the wire added in the loader. */
_search?: string;
}
const allSkills: Skill[] = skills as Skill[];
const allSkills: Skill[] = [];
interface IndexMeta {
extractedAt?: string;
@ -32,7 +35,7 @@ interface IndexMeta {
externalSource?: string;
bySource?: Record<string, number>;
}
const indexMeta: IndexMeta = meta as IndexMeta;
const indexMeta: IndexMeta = {};
function formatRelativeTime(iso?: string): string | null {
if (!iso) return null;
@ -398,8 +401,43 @@ function StatCard({ value, label, color }: { value: number; label: string; color
const PAGE_SIZE = 60;
// Routes Docusaurus serves the static API JSON from. `baseUrl` is `/docs/`,
// `static/api/` ends up at `/docs/api/`. Hardcoding here is fine because the
// same `baseUrl` is enforced repo-wide; if it ever changes, this is the only
// place that needs to follow.
const SKILLS_URL = "/docs/api/skills.json";
const META_URL = "/docs/api/skills-meta.json";
function buildSearchHaystack(s: Skill): string {
// Pre-compute the lowercase blob the search filter scans. Done once at
// load time instead of per-keystroke per-skill. With 50k+ skills the
// per-keystroke variant was unusably slow.
return [
s.name,
s.description,
s.overview,
s.categoryLabel,
s.author,
...(s.tags || []),
]
.filter(Boolean)
.join(" ")
.toLowerCase();
}
export default function SkillsDashboard() {
// Lazy-loaded data. Was bundled into the JS chunk (~22 MB at 50k skills,
// which made the initial page load unusable on mobile). Now fetched on
// mount from the same CDN that serves the docs.
const [data, setData] = useState<{ skills: Skill[]; meta: IndexMeta } | null>(null);
const [loadError, setLoadError] = useState<string | null>(null);
const [search, setSearch] = useState("");
// Debounced copy of `search` — used by the filter. Without the debounce,
// typing into the search box ran .filter() over the whole catalog on
// every keystroke, which on a 50k-item list felt like the page had
// hung. 150ms gives a snappy feel without lagging behind the user.
const [debouncedSearch, setDebouncedSearch] = useState("");
const [sourceFilter, setSourceFilter] = useState("all");
const [categoryFilter, setCategoryFilter] = useState("all");
const [expandedCard, setExpandedCard] = useState<string | null>(null);
@ -408,6 +446,42 @@ export default function SkillsDashboard() {
const searchRef = useRef<HTMLInputElement>(null);
const gridRef = useRef<HTMLDivElement>(null);
useEffect(() => {
let cancelled = false;
(async () => {
try {
const [sk, mt] = await Promise.all([
fetch(SKILLS_URL).then((r) => {
if (!r.ok) throw new Error(`skills.json HTTP ${r.status}`);
return r.json();
}),
fetch(META_URL).then((r) => (r.ok ? r.json() : {})).catch(() => ({})),
]);
if (cancelled) return;
const skillsArr = Array.isArray(sk) ? (sk as Skill[]) : [];
// Stamp the precomputed search haystack onto each row.
for (const s of skillsArr) s._search = buildSearchHaystack(s);
setData({ skills: skillsArr, meta: mt || {} });
} catch (err) {
if (cancelled) return;
setLoadError(err instanceof Error ? err.message : String(err));
}
})();
return () => {
cancelled = true;
};
}, []);
// Debounce the search input — 150ms feels instant while preventing the
// filter from running on every individual keystroke.
useEffect(() => {
const t = setTimeout(() => setDebouncedSearch(search), 150);
return () => clearTimeout(t);
}, [search]);
const allSkillsLocal: Skill[] = data?.skills ?? [];
const indexMetaLocal: IndexMeta = data?.meta ?? indexMeta;
useEffect(() => {
const handler = (e: KeyboardEvent) => {
if (e.key === "/" && document.activeElement?.tagName !== "INPUT") {
@ -424,15 +498,15 @@ export default function SkillsDashboard() {
}, []);
const sources = useMemo(() => {
const set = new Set(allSkills.map((s) => s.source));
const set = new Set(allSkillsLocal.map((s) => s.source));
return SOURCE_ORDER.filter((s) => s === "all" || set.has(s));
}, []);
const categoryEntries = useMemo(() => {
const pool =
sourceFilter === "all"
? allSkills
: allSkills.filter((s) => s.source === sourceFilter);
? allSkillsLocal
: allSkillsLocal.filter((s) => s.source === sourceFilter);
const map = new Map<string, { label: string; count: number }>();
for (const s of pool) {
const key = s.category || "uncategorized";
@ -452,24 +526,22 @@ export default function SkillsDashboard() {
}, [sourceFilter]);
const filtered = useMemo(() => {
const q = search.toLowerCase().trim();
return allSkills.filter((s) => {
const q = debouncedSearch.toLowerCase().trim();
return allSkillsLocal.filter((s) => {
if (sourceFilter !== "all" && s.source !== sourceFilter) return false;
if (categoryFilter !== "all" && s.category !== categoryFilter) return false;
if (q) {
const haystack = [s.name, s.description, s.overview, s.categoryLabel, s.author, ...(s.tags || [])]
.join(" ")
.toLowerCase();
return haystack.includes(q);
// _search is pre-built in the load effect — single .includes() per row.
return (s._search || "").includes(q);
}
return true;
});
}, [search, sourceFilter, categoryFilter]);
}, [debouncedSearch, sourceFilter, categoryFilter, allSkillsLocal]);
useEffect(() => {
setVisibleCount(PAGE_SIZE);
setExpandedCard(null);
}, [search, sourceFilter, categoryFilter]);
}, [debouncedSearch, sourceFilter, categoryFilter]);
const visible = filtered.slice(0, visibleCount);
const hasMore = visibleCount < filtered.length;
@ -512,15 +584,22 @@ export default function SkillsDashboard() {
<h1 className={styles.heroTitle}>Skills Hub</h1>
<p className={styles.heroSub}>
Discover, search, and install from{" "}
<strong className={styles.heroAccent}>{allSkills.length}</strong> skills
across {sources.length - 1} registries
<strong className={styles.heroAccent}>
{data ? allSkillsLocal.length.toLocaleString() : "…"}
</strong>{" "}
skills across {sources.length - 1} registries
{loadError && (
<span style={{ color: "#f87171", marginLeft: 8 }}>
· failed to load catalog ({loadError})
</span>
)}
</p>
{(indexMeta?.indexGeneratedAt || indexMeta?.extractedAt) && (
{(indexMetaLocal?.indexGeneratedAt || indexMetaLocal?.extractedAt) && (
<p className={styles.heroSub} style={{ fontSize: "0.85rem", opacity: 0.75 }}>
Catalog refreshed{" "}
<span title={indexMeta.indexGeneratedAt || indexMeta.extractedAt}>
<span title={indexMetaLocal.indexGeneratedAt || indexMetaLocal.extractedAt}>
{formatRelativeTime(
indexMeta.indexGeneratedAt || indexMeta.extractedAt,
indexMetaLocal.indexGeneratedAt || indexMetaLocal.extractedAt,
) || "recently"}
</span>
{" "}· auto-rebuilt twice daily
@ -529,18 +608,18 @@ export default function SkillsDashboard() {
<div className={styles.statsRow}>
<StatCard
value={allSkills.filter((s) => s.source === "built-in").length}
value={allSkillsLocal.filter((s) => s.source === "built-in").length}
label="Built-in"
color="#4ade80"
/>
<StatCard
value={allSkills.filter((s) => s.source === "optional").length}
value={allSkillsLocal.filter((s) => s.source === "optional").length}
label="Optional"
color="#fbbf24"
/>
<StatCard
value={
allSkills.filter(
allSkillsLocal.filter(
(s) => s.source !== "built-in" && s.source !== "optional"
).length
}
@ -548,7 +627,7 @@ export default function SkillsDashboard() {
color="#60a5fa"
/>
<StatCard
value={new Set(allSkills.map((s) => s.category)).size}
value={new Set(allSkillsLocal.map((s) => s.category)).size}
label="Categories"
color="#a78bfa"
/>
@ -592,8 +671,8 @@ export default function SkillsDashboard() {
const conf = SOURCE_CONFIG[src];
const count =
src === "all"
? allSkills.length
: allSkills.filter((s) => s.source === src).length;
? allSkillsLocal.length
: allSkillsLocal.filter((s) => s.source === src).length;
return (
<button
key={src}